Hello there,
in my case the prompt_tokens are about 15-20x higher then completion_tokens. I think in this case a fine-tuned model makes sense.
I created a list of 50 examples of conversations. Here is an example of 1/50. Of course the system prompt is always the same.
{"messages": [{"role": "system", "content": "you are an assistant that always generates JSON. you always return just the JSON with no additional description or context.\n\nI will give you a text with details about a watch. if you can not find any information in the given text, remove the key from the json response. In the values you find the possible values of each key.\n\n###\n\n{\n \"name\": \"get_watch_data\",\n \"description\": \"Extract technical data from a text about a wristwatch.\",\n \"parameters\": {\n \"type\": \"object\",\n \"properties\": {\n \"brand\": {\n \"type\": \"string\",\n \"description\": \"brand of the watch\"\n },\n \"model\": {\n \"type\": \"string\",\n \"description\": \"model of the watch\"\n },\n \"reference\": {\n \"type\": \"string\",\n \"description\": \"reference number of the watch\"\n },\n \"price\": {\n \"type\": \"integer\",\n \"description\": \"price of the watch in euro\"\n },\n \"year\": {\n \"type\": \"integer\",\n \"description\": \"year of manufacture of the watch in format YYYY\"\n },\n \"condition\": {\n \"type\": \"string\",\n \"description\": \"condition of the watch\",\n \"values\": \"new, used\"\n },\n \"boxed\": {\n \"type\": \"boolean\",\n \"description\": \"whether the watch comes with the original box\"\n },\n \"papers\": {\n \"type\": \"boolean\",\n \"description\": \"whether the watch comes with the original papers\"\n },\n \"seller_type\": {\n \"type\": \"string\",\n \"description\": \"type of seller\",\n \"values\": \"private, commercial\"\n },\n \"watch_type\": {\n \"type\": \"string\",\n \"description\": \"type of watch\",\n \"values\": \"men, women, unisex\"\n },\n \"movement\": {\n \"type\": \"string\",\n \"description\": \"type of movement\",\n \"values\": \"automatic, manual, quartz, solar\"\n },\n \"movement_function\": {\n \"type\": \"string\",\n \"description\": \"functions of the movement. multiple values are separated by comma.\",\n \"values\": \"chronograph, date, split_second_chronograph, perpetual_calendar, flyback_function, gmt_second_time_zone, annual_indicator, annual_calendar, minute_repeater, month_indicator, moon_phase, panorama_date, chiming_mechanism, jumping_hour, tachymeter, tourbillon, four_year_calendar, day_of_the_week_indicator, equation_of_time\"\n },\n \"dial\": {\n \"type\": \"object\",\n \"properties\": {\n \"format\": {\n \"type\": \"string\",\n \"description\": \"format of the dial\",\n \"values\": \"digital, arabic, no_digits, roman\"\n },\n \"color\": {\n \"type\": \"string\",\n \"description\": \"color of the dial\",\n \"values\": \"blue, bordeaux, brown, bronze, champagne, yellow, gold, gold_solid, gray, green, orange, mother_of_pearl, pink, red, black, silver, silver_solid, transparent, violet, white\"\n }\n }\n },\n \"house\": {\n \"type\": \"object\",\n \"properties\": {\n \"material\": {\n \"type\": \"string\",\n \"description\": \"material of the watch case\",\n \"values\": \"gold, platinum, stainless_steel, titanium, ceramic, carbon, aluminum, brass, bronze, palladium, rhodium, silver, tungsten, white_gold, yellow_gold, rose_gold, red_gold, pink_gold, black_gold, blue_gold, green_gold, purple_gold, orange_gold, brown_gold, gray_gold, black_pvd, blue_pvd, green_pvd, purple_pvd, orange_pvd, brown_pvd, gray_pvd, black_dlc, blue_dlc, green_dlc, purple_dlc, orange_dlc, brown_dlc, gray_dlc\"\n },\n \"bezel\": {\n \"type\": \"string\",\n \"description\": \"material of the watch bezel\",\n \"values\": \"gold, platinum, stainless_steel, titanium, ceramic, carbon, aluminum, brass, bronze, palladium, rhodium, silver, tungsten, white_gold, yellow_gold, rose_gold, red_gold, pink_gold, black_gold, blue_gold, green_gold, purple_gold, orange_gold, brown_gold, gray_gold, black_pvd, blue_pvd, green_pvd, purple_pvd, orange_pvd, brown_pvd, gray_pvd, black_dlc, blue_dlc, green_dlc, purple_dlc, orange_dlc, brown_dlc, gray_dlc\"\n },\n \"glass\": {\n \"type\": \"string\",\n \"description\": \"material of the watch glass\",\n \"values\": \"sapphire, mineral, plexiglass, hesalite, acrylic, plastic, glass, crystal\"\n },\n \"waterproof\": {\n \"type\": \"integer\",\n \"description\": \"water resistance of the watch in meters\"\n },\n \"diameter\": {\n \"type\": \"integer\",\n \"description\": \"diameter of the watch case in millimeters\"\n },\n \"height\": {\n \"type\": \"integer\",\n \"description\": \"height of the watch case in millimeters\"\n }\n }\n },\n \"bracelet\": {\n \"type\": \"object\",\n \"properties\": {\n \"material\": {\n \"type\": \"string\",\n \"description\": \"material of the watch bracelet\",\n \"values\": \"leather, textile, rubber, metal, ceramic, titanium, gold, platinum, stainless_steel, carbon, aluminum, brass, bronze, palladium, rhodium, silver, tungsten, white_gold, yellow_gold, rose_gold, red_gold, pink_gold, black_gold, blue_gold, green_gold, purple_gold, orange_gold, brown_gold, gray_gold, black_pvd, blue_pvd, green_pvd, purple_pvd, orange_pvd, brown_pvd, gray_pvd, black_dlc, blue_dlc, green_dlc, purple_dlc, orange_dlc, brown_dlc, gray_dlc\"\n },\n \"color\": {\n \"type\": \"string\",\n \"description\": \"color of the watch bracelet\",\n \"values\": \"black, blue, brown, bronze, champagne, yellow, gold, gold_solid, gray, green, orange, pink, red, silver, silver_solid, transparent, violet, white, stainless_steel\"\n },\n \"width\": {\n \"type\": \"integer\",\n \"description\": \"width of the watch bracelet in millimeters\"\n }\n }\n },\n \"clasp\": {\n \"type\": \"object\",\n \"properties\": {\n \"material\": {\n \"type\": \"string\",\n \"description\": \"material of the watch clasp\",\n \"values\": \"gold, platinum, stainless_steel, titanium, ceramic, carbon, aluminum, brass, bronze, palladium, rhodium, silver, tungsten, white_gold, yellow_gold, rose_gold, red_gold, pink_gold, black_gold, blue_gold, green_gold, purple_gold, orange_gold, brown_gold, gray_gold, black_pvd, blue_pvd, green_pvd, purple_pvd, orange_pvd, brown_pvd, gray_pvd, black_dlc, blue_dlc, green_dlc, purple_dlc, orange_dlc, brown_dlc, gray_dlc\"\n },\n \"type\": {\n \"type\": \"string\",\n \"description\": \"type of the watch clasp\",\n \"values\": \"buckle, folding_clasp, deployant_clasp, tang_buckle, pin_buckle, butterfly_clasp, push_button_clasp, safety_clasp, hook_buckle, hook_clasp, lobster_clasp, snap_clasp, slide_clasp, twist_lock_clasp, velcro_clasp, magnetic_clasp, push_lock_clasp, screw_down_clasp, spring_bar_clasp, tang_buckle_clasp, tongue_buckle_clasp, twist_lock_buckle_clasp\"\n }\n }\n }\n },\n \"required\": [\n \"model\",\n \"price\",\n \"brand\"\n ]\n }\n}\n"}, {"role": "user", "content": "BREITLING Colt Automatik 41, Stahlband, graues Zifferblatt, Ref. A17313101F1A1\nNochmal Hallo Zusammen.\n\nNoch so eine mit Stahll\u00fcnette die mich verlassen darf: Breitling Colt Automatik 41 mm. Sie stammt aus November 2019 und hat ein graues Zifferblatt mit Sonnernchliff was mich damals ziemlich geflasht hatte und das es so aufger\u00e4umt ist, finde ich tats\u00e4chlich immer noch sch\u00f6ner als andere Modellschwestern aus der Colt Serie die ja bekanntlich eingestellt wurde.\n\nDie Uhr wurde getragen, wenn auch selten, hat wenige Swirls die bei so viel Stahl \u00fcberhaupt nicht zu verhindern sind.\n\nDies ist ein Privatverkauf unter Ausschluss der Gew\u00e4hrleistung bzw. Sachm\u00e4ngelhaftung. Ein Umtausch und die R\u00fcckgabe ist ebenfalls ausgeschlossen. Ich bin nicht an einem Tausch interessiert und m\u00f6chte auch nichts ins Zahlung nehmen.\n\nIch stelle mir preislich sehr faire \u20ac 2.250,- bei Abholung und Zahlung hier in Bonn vor. Ein Versand per Wertkurier bei Zahlung vorab und \u00dcbernahme der Kosten ist m\u00f6glich. Ich habe einige, ausschlie\u00dflich gute, Bewertungen unter \"User bewerten User\""}, {"role": "assistant", "content": "{\n \"brand\": \"BREITLING\",\n \"model\": \"Colt Automatik 41\",\n \"reference\": \"A17313101F1A1\",\n \"price\": 2250,\n \"year\": 2019,\n \"condition\": \"used\",\n \"seller_type\": \"private\",\n \"dial\": {\n \"color\": \"gray\"\n },\n \"house\": {\n \"material\": \"stainless_steel\"\n }\n}"}]}
I tried different epoch 3, 10 and 32 but I only get the results if I use the system-prompt like before. I thought I can reduce the input tokens with the trained model?
Have I made a mistake?